Yesterday I noticed for the first time that Claude identified and rejected AI-written content that it found during web search:
The search results are poisoned
Searching this returns confident-looking lists of specialist results. [...] they don't exist. [...] It's AI-generated SEO copy from a supplement company, and it's the top result. [...]
Yelp's "low fodmap San Francisco" list is just keyword matching.
This is the first time I've felt like "dead internet theory" and the content collapse is starting to hit a tipping point.
It's curious to see how it recognized the poisoned content. AFAICT, it wasn't from identifying any special fingerprint in the content (e.g. statistical character recognition or detecting hidden characters) but from it researching its result ("I chased them down and they don't exist."). This is a different method than Anthropic's announcement about embedded watermarks as part of the EU AI Act’s Code of Practice on Transparency of AI-Generated Content.
But I suspect the results will be the same: AI-written content will begin to be flagged or deprioritized in agentic workflows and search.
I'm not mad about that.